tonbistudio/turboquant-pytorch

View on GitHub ↗ ★ 1,048 Python

From-scratch PyTorch implementation of Google's TurboQuant (ICLR 2026) for LLM KV cache compression. 5x compression at 3-bit with 99.5% attention fidelity.

3 mentions on Hacker News by 3 people: 1 story, 2 comments · first 2026-03-25 · latest 2026-04-03

Mentions per year 3 mentions · busiest year 2026 (3)

20082026

All Stories Comments Newest Top Oldest

▲ 2
10-19x faster alternative to TurboQuant via Clifford rotors (44x fewer params)
story by hochmartinez · 2026-04-03 · 0 comments